Abstract:

How can we train a robot to generalize to diverse environments? This question underscores the holy grail of robot learning research because it is difficult to supervise an agent for all possible situations it can encounter in the future. We posit that the only way to guarantee such a generalization is to continually learn and adapt to new situations. This adaptation has to be rapid and occur online at a time scale of fractions of a second, which implies that we have no time to carry out multiple experiments in the physical world to estimate system parameters. In this talk, I will present our early efforts in this direction by focusing on the case study of legged locomotion. I will talk in-depth about our two recent papers:

(a) Rapid Motor Adaptation (RSS 2021): RMA allows a quadruped robot to adapt online in half a second to changing terrains. The approach is purely learning-based without making use of any control stack or demonstrations or predefined leg swing motions. The robot is trained in simulation and then rapidly adapts when introduced to the real world.

(b) Energy minimization leads to the emergence of gaits (CoRL 2021): Instead of hand-designing the leg swing motions, we show that natural gaits emerge if we minimize energy consumed within the RMA framework. This is consistent with various biomechanics studies in the last century on animals like horses, dogs, etc. We find that our robot is in the category of a horse by Froude number analysis and show the emergence of horse-like gaits from scratch as the speed changes — walk at low speed, then trot at med speed, and then gallop/canter at high speeds.

Bio:

Deepak Pathak is a faculty in the School of Computer Science at Carnegie Mellon University. He received his Ph.D. from UC Berkeley and his research spans computer vision, machine learning, and robotics. He is a recipient of the faculty awards from Google, Sony, GoodAI, and graduate fellowship awards from Facebook, NVIDIA, Snapchat. His research has been featured in popular press outlets, including The Economist, The Wall Street Journal, Quanta Magazine, Washington Post, CNET, Wired, and MIT Technology Review among others. Deepak received his Bachelor’s from IIT Kanpur with a Gold Medal in Computer Science. He co-founded VisageMap Inc. later acquired by FaceFirst Inc.

Homepage: https://www.cs.cmu.edu/~dpathak/

Sponsored in part by:   Facebook Reality Labs Pittsburgh